Trump versus Clinton – Twitter Communication
During the US Primaries
Jennifer Fromm, Stefanie Melzer, Bj¨orn Ross (0000-0003-2717-3705), and
Stefan Stieglitz (0000-0002-4366-1840)
University of Duisburg-Essen, Duisburg, Germany,
Abstract. When Donald Trump won the Republican nomination and
subsequently beat Hillary Clinton in the presidential elections, his success
came as a surprise to most observers. This research contributes to under-
standing the dynamics of this unusual campaign, in which social media
played a prominent role. We collected 6,099 tweets by both nominees
during the presidential primaries and identiﬁed the 21 most frequently
discussed issues through computer-assisted content analysis. Secondly,
we used time series analysis to investigate whether the candidates inﬂu-
enced each other’s political agendas. Most tweets by the candidates were
found not to be about policy but about parties, other politicians, and
the media. Of the political issues that were discussed, the most promi-
nent ones were employment, family, minorities and terrorism. For tweets
about minorities, we found possible evidence of agenda setting. We con-
clude that social media are mainly being used to reach out to supporters,
instead of interacting with the opponent.
Keywords: microblog analysis, Twitter, agenda setting, inter-policy
agenda setting, content analysis, time series analysis
The results of the United States presidential primaries in 2016 were highly unex-
pected. Donald Trump, a businessman with no prior experience in political oﬃce,
was ﬁrst seen as an outsider with no real chance at winning the nomination. “If
Trump is nominated, then everything we think we know about presidential nom-
inations is wrong”, researchers at the University of Virginia Center for Politics
said on their blog in August 2015 . Trump was nominated and ultimately
elected president, so what do we know?
The 2016 primaries were exceptionally polarizing and emotional. Donald
Trump insulted the other Republican candidates, the Democratic candidates,
politicians from all over the world, the media, women and ethnic groups .
Researchers have argued that Trump oﬀers the masculine image of a “tough
guy”  and anti-politician who channels dissatisfaction . It has also been
argued in the media that Trump has been successful at using social media to
rally his supporters . However, conclusive scientiﬁc evidence for this assertion
is so far missing.
It is clear, however, that both Trump and Clinton, or their respective cam-
paign staﬀ, are proliﬁc and inﬂuential Twitter users. Clinton has more than 8
million followers and has written about 7,000 short messages, or tweets, since
2013. Donald Trump has more than 10 million followers and has written more
than 32,000 tweets since the creation of his account in 2009 (as of August 6,
2016). Both candidates are evidently able to reach millions of people on Twit-
ter. Their tweets reﬂect their standpoints on political issues and their style of
campaigning. In an analysis by the Washington Post of more than 6000 tweets
posted by Donald Trump between June and December 2015, 11 % were found to
be insulting .
The last decade has seen the emergence of social media and its acceptance as
a new useful tool for researchers to examine a wide array of phenomena in do-
mains such as politics and business [6–8]. In particular, it has helped researchers
understand the dynamics of electoral campaigns [9–12]. It has been studied how
social media data can be used to predict voter behaviour , how new tech-
nologies have shaped political campaigning  and which role social media
appearances play in campaigns . We contribute to this body of research by
examining which issues were discussed on Twitter by the nominees in the run-up
to the 2016 election, and how political issues were discussed.
To do so, we draw on agenda setting theory. This theory distinguishes three
diﬀerent agendas: the public agenda, the media agenda, and the policy agenda.
For example, in this framework, the communication by the candidates on Twit-
ter could be viewed as a reﬂection of the future policy agenda. Moreover, since
communication on Twitter is generally public, candidates could inﬂuence each
other in their communication and respond to one another. Therefore, this re-
search addresses the following questions:
1. What is the nature of the Twitter communication by the main presidential
candidates during the 2016 US primaries?
2. To what extent does agenda setting take place on Twitter between these
These research questions are addressed using a combination of content analy-
sis and time series analysis. Content analysis is used to identify the most promi-
nent topics discussed during the primaries by both eventual nominees (i.e. the
ﬁrst research question). Time series analysis is then used to address the second
research question by examining the interrelations between the topic mentions
over time. Prior research has investigated agenda setting on Twitter [16,17], but
the application of this combination of methods and a social media data source
to study inter-candidate agenda setting is new.
In the remainder of the article, we give a theoretical overview and introduce
the present literature on agenda setting. We especially focus on the policy agenda
and Twitter. Our hypotheses are derived from the literature. The third section
describes the methods used, and section four presents the empirical results. Sec-
tion ﬁve contains a discussion of the results and our conclusions. We ﬁnally
consider the limitations of this study and its implications for future research.
2 Related Work
Agenda setting theory concerns the idea that people do not only learn about
a certain topic by media consumption, they are also learning about its impor-
tance by evaluating the place and space of this speciﬁc topic . In the words
of Cohen [19, p. 13], media don’t tell people “what to think”, but “[. . . ] what
to think about”. According to agenda setting theory, there are three diﬀerent
agendas: The public agenda, the media agenda and the policy agenda. These
agendas are interrelated and inﬂuence each other. Personal experience, interper-
sonal communication and the “real world” have been found to inﬂuence all three
agendas . In the context of this study, the policy agenda is the most relevant.
2.1 The Policy Agenda
The policy agenda describes the actions taken by the government. The agendas of
political parties, the bureaucracy, the President, the Committees and the Lower
and Upper House belong to the policy agenda . In the policy agenda, issues
are of great importance. It is crucial for the candidates to ﬁnd the majority-
eﬃcient position for a certain issue to convince voters . Candidates will give
more salience to issues for which they get broader support from voters .
Studies about policy agenda setting pay attention to the dynamics in the political
system and answer the question of how a new idea, policy or problem is accepted
in the political system .
The relationship between the media agenda and the policy agenda is recip-
rocal, that is, both agendas inﬂuence each other: Policy makers are not inde-
pendent of the media – and the media are rarely independent of members of
political institutions . On the one hand, the president has been shown to
inﬂuence the media agenda on foreign policy issues. In particular, issues with
lower salience concerning foreign policy are most likely to be taken up by the
media . Moreover, the president has been found to be cited regularly during
news about a press conference . On the other hand, the media also inﬂuence
the policy agenda by giving more attention to certain issues than to other is-
sues. Media attention is often seen as “an agent of change” [25, p. 110] that has
a stabilizing power for the policy-making process.
Politicians can take up issues from others and voice their opinions about
these issues to distinguish themselves from other politicians. Soroka  calls
these dynamics within political institutions inter-policy agenda setting. Stud-
ies regarding these dynamics found such eﬀects between the President of the
United States and Congress, with the former setting the agenda of the latter in
most cases. Only for the issue of international aﬀairs, Congress sets the Pres-
ident’s agenda . Moreover, the agenda setting process between candidates
during election times has gained the attention of researchers during the last
years. Inter-candidate agenda setting takes place for both issue salience and at-
tribute salience relationships, but the results are stronger for attribute salience.
The authors deﬁned salience “by the frequency of issue and attribute mentions
within campaign messages” .
2.2 Agenda Setting and Twitter
Campaign messages today are not only disseminated through traditional media,
but also through social media such as the microblogging service Twitter. One
can use the service to publish a status (tweet) with a maximum length of 140
characters. Users can also follow each other, that is, be notiﬁed when someone
else publishes a tweet. Since its launch in 2006, Twitter has become a very
popular medium in the US, with 66 million monthly active users in June 2016
Even in times of a more fragmented media landscape, agenda setting takes
place , but in a diﬀerent way: Traditional media have become less powerful
in the agenda-setting process, as their former power is now divided between
traditional media and citizen media . Since the election of Obama in 2008,
Twitter has been widely used by politicians, especially during election periods
. Obama used Facebook and Twitter to collect donations and to connect to
the community . Politicians are known to have an inﬂuential role on Twitter
in terms of retweets .
Research regarding agenda setting and Twitter has particularly focused on
possible inter-media agenda setting eﬀects. In a political context, it has been
found that traditional media such as newspapers or television have a “symbi-
otic relationship” [16, p. 374] with the Twitter feeds of candidates and political
parties for certain issues such as employment and health care during election
times [16, 35].
Another research area concerning this topic is network agenda setting. This
theory examines agenda setting eﬀects between the media and public agenda
from a network perspective. Using Big Data analysis, this theory has been con-
ﬁrmed for the 2012 election, when Mitt Romney was the Republican candidate
who competed with Barack Obama. The authors found that the network issue
agendas of the candidates’ supporters correlated positively with the network is-
sue agendas of certain media channels, in particular of horizontal media such as
talk shows and cable news . Especially young Americans who are part of the
Twitter network also search for information on political issues and express their
In summary, Twitter is a useful and promising tool for studying agenda
setting, but inter-policy agenda setting has not yet been explored in this context.
We therefore test the following hypotheses on Twitter data.
Inter-policy agenda setting can take place between parties. Vliegenthart et al.
 examined the dynamics of the policy agenda in Belgium in a long-term study.
They found that the parties inﬂuence each other and are more likely to take up
the issues of other parties in parliament if they are from the same language
community. Moreover, governing parties have more agenda setting power than
other parties. Hillary Clinton belongs to the governing party in the US, was part
of it as United States Secretary, and is also supported by the President. We
therefore hypothesize that she will lead the agenda of Donald Trump:
H1 The issues mentioned in Hillary Clinton’s tweets will predict the issues
mentioned in Donald Trump’s tweets.
Tedesco  found inter-candidate agenda setting eﬀects in the candidate
and campaign press releases of the Democratic candidates during the primaries
in 2004. The author assumes that the agenda of the Democratic candidate John
Kerry was set by the shared agenda of the other Democratic candidates. His
opponent Howard Dean gained ﬁnancial support and large media attention early,
but he did not lead the other candidates’ agendas. According to the author, Dean
was not able to take advantage of the media attention. Furthermore, Vliegenthart
et al.  ﬁnd that extreme-right parties also have an agenda-setting power.
In 2016, Donald Trump received an exceptional amount of media attention,
which he might be able to take advantage of. He has more Twitter followers
than Clinton and makes polarizing statements that other candidates may have
no choice but to react to. Thus, we hypothesize:
H2 The issues mentioned in Donald Trump’s tweets will predict the issues
mentioned in Hillary Clinton’s tweets.
In summary, there are good reasons to believe that the candidates should
inﬂuence each other’s agenda. Examining whether this is indeed the case helps
understand how Twitter was used by the candidates before the election.
To address the research questions and test the hypotheses, we chose a quantita-
tive research design. A large dataset of relevant tweets was collected. Afterwards
we conducted a content analysis to identify the most relevant political topics
during the US primaries 2016. In a time series analysis, we focused on these
identiﬁed topics and examined if any agenda setting eﬀects occurred.
As the present analysis focuses on the two main presidential primary candidates,
only tweets and retweets by Hillary Clinton and Donald Trump (@HillaryClinton
and @realDonaldTrump) were collected via the GET statuses/user timeline
endpoint of the Twitter API. This means we queried the API for all tweets and
retweets sent from Donald Trump’s and Hillary Clinton’s accounts. Contrary to
other research on Twitter, we did not search for tweets containing hashtags or
keywords related to the candidates, as our research focused on the communi-
cation between the candidates and not on the communication of other Twitter
users about them. Every tweet posted between November 15, 2015 and June
4, 2016 was collected, because the primaries took place within this timespan.
Hillary Clinton occasionally tweets in Spanish. These tweets were excluded from
further analysis. The cleaned dataset contained 6,099 tweets. Of these, 3,056
were posted by Donald Trump and 3,043 by Hillary Clinton, so we were able to
analyze a similar amount of tweets by each candidate.
3.2 Content Analysis
To address the ﬁrst research question – that is, to examine the communication
on Twitter by both candidates during the campaign –, we conducted a content
analysis of the tweets and determined the most salient topics in each candidate’s
messages. As the present dataset is too large to analyze them manually, the
content analyis was conducted in a computer-assisted way.
The computer-assisted qualitative coding program QDA Miner and its text
mining component WordStat were used to analyze the most frequent topics of
both candidates. WordStat oﬀers a few ready-to-use dictionaries. After testing
those dictionaries on a small sample dataset, none of those dictionaries were
found to be suitable for this research context. As political topics diﬀer across
elections, we used WordStat to develop a suitable dictionary ourselves.
The literature search revealed that Conway et al.  also conducted a
computer-assisted content analysis to analyze agenda setting during US pri-
maries. They used a dictionary including 21 political topics. These categories
were used as a starting point for our own dictionary. First of all, all tweets were
entered into QDA Miner and word frequencies were analyzed in WordStat. The
most frequent words were classiﬁed and put into the dictionary categories based
on the work of Conway et al. . As expected, topics during these primaries
diﬀered from previous primaries, so not all categories by Conway et al. 
were used and some new categories were added. As several words have diﬀerent
meanings, we resolved uncertainties by using the Keyword-in-Context tool. This
shows all tweets including the word in question, so it is easier to decide into
which category a word belongs.
After adding a decent amount of words, we conducted a ﬁrst content analysis
based on our own dictionary. WordStat was conﬁgured to use Porter Stemming
and a built-in English exclusion list. Porter Stemming removes common English
preﬁxes and suﬃxes before the categorization process. The exclusion list con-
tained English stop words which provide no further meaning to a text (e.g. and,
or, the). The built-in list was manually extended with Twitter-related stop words
such as RT (the abbreviation of retweet). During the categorization process, the
dictionary recognizes the beginnings and endings of words by identifying space
characters. If a tweet contained one or more words included in the dictionary, it
was assigned to all matching categories. Results returned a list of leftover words
which were not included in the dictionary yet. We classiﬁed the most frequent
leftover words to make sure that our dictionary contains all words occurring in
more than one percent of all tweets.
Our ﬁnal dictionary includes the following 21 categories: employment,envi-
ronment,guns,health care,military and defense,terrorism,slogans,media,fam-
ily,rights,meetings,thank-you messages,campaign funding,parties and politi-
cians,caucus,foreign politics,education,economics,justice,Trump family and
To validate our dictionary, we calculated recall and precision for each cate-
gory. Therefore, 60 random tweets were coded manually ﬁrst. Afterwards, the
same tweets were coded by WordStat, using the developed dictionary. We calcu-
lated recall and precision values for each category. Scores for both measures can
range from 0 to 1.0, where 1.0 would be the ideal result. A recall of 1.0 means
that all tweets belonging to a speciﬁc category were labelled as belonging to this
category by the dictionary, but says nothing about how many other tweets were
labelled incorrectly as belonging to this category. A precision score of 1.0 means
that every tweet labelled as belonging to a speciﬁc category indeed belongs to
this category, but says nothing about the number of tweets that also belong to
this category but were labelled incorrectly . Most categories reached high
values in recall and precision, but there are a few exceptions, for example the
category justice (R= 0.50, P = 1). In sum, the dictionary reached an average
recall of 0.84 and an average precision of 0.97 over all categories.
3.3 Time Series Analysis
To address the second research question and study the interrelations between
the agendas of the two candidates, we used time series analysis. The content
analysis served as a preprocessing step for the time series analysis. The agenda
setting hypotheses were tested on political topics which were discussed frequently
by both candidates. To identify these topics, the results of the content analysis
During content analysis, the political topics family,employment,minorities
and terrorism were identiﬁed as the ones most frequently discussed by both
candidates. Therefore the dataset for time series analysis contained only tweets
which were classiﬁed into at least one of these categories. For every candidate
and topic, a time series was created resulting in eight diﬀerent time series. All of
these contain date and number of tweets related to a speciﬁc topic as variables.
As Twitter is a very fast-paced medium, the number of tweets related to a
speciﬁc topic was calculated for every day. The following method for analysing
time series data was also adapted from Conway et al. , as their research took
place in a similar context.
The statistic software SPSS was used to analyze relationships between Donald
Trump’s and Hillary Clinton’s time series. At ﬁrst, all time series were tested
for auto-correlations, which are correlations within a time series. Next, all time
series were examined for linear trends by using curve estimation. For every time
series with signiﬁcant linear trends SPSS automatically calculated a de-trended
version of the original time series as a new variable. Finally, cross-correlations for
every topic were calculated. Every time Donald Trump’s time series was entered
as the ﬁrst variable and Hillary Clinton’s as the second variable. In case a linear
trend was found in the previous step, the de-trended time series was entered
In the following section results from content analysis and time series analysis are
presented. At the end, we summarize shortly which hypotheses are supported
by our results.
4.1 Content Analysis
The topic distributions of Donald Trump’s and Hillary Clinton’s tweets are
shown in Table 1. It stands out that both candidates tweeted mostly about
their media appearances, parties, and other politicians. Topics regarding politi-
cal issues, for example, employment,healthcare or human rights were discussed
much less often. Employment was the most frequently discussed political issue
by Donald Trump, but occurred only in 4 % of his tweets. This means that all
other political issues were covered even less frequently. Hillary Clinton’s topic
distribution looks quite similar, but there are some diﬀerences. Her most fre-
quently discussed political issue family occurred in 16 % of her tweets. So both
candidates had a diﬀerent favourite political issue and Hillary Clinton tweeted
more often about her favourite issue family than Donald Trump tweeted about
employment. Whereas Donald Trump discussed all political issues rarely, Hillary
Clinton has one clear main issue, but tweeted about all other political issues
as seldom as Trump. Minorities as her second political issue, for example, ac-
counted only for 6 % of her tweets.
The main goal of the content analysis was to identify political topics appro-
priate for the subsequent time series analysis. These topics should occur quite
often in both candidates’ tweets, so that enough data for valid results are avail-
able. We decided to choose the most frequently discussed political topics from
Trump’s topic distribution and checked if they also occur often enough in Clin-
ton’s tweets. As a result, the political topics employment,family,terrorism and
minorities were chosen for further analysis.
4.2 Time Series Analysis
As described in the method section every time series was tested for auto-correlations
and linear trends. In the following time series signiﬁcant auto-correlations were
found: Trump family (lag 1, lag 2), Trump terrorism (lag 1, 2, 3 and 14), Clin-
ton minorities (lag 1) and Clinton family (lag 14). Signiﬁcant linear trends were
found in these time series: Trump minorities (R2=.03, p < .05), Trump terror-
ism (R2=.05, p < .01), Clinton family (R2=.02, p =.05) and Clinton terrorism
(R2=.03, p =.01).
In the next step, cross-correlation coeﬃcients were calculated for all four
topics. Table 2 shows that for the topic minorities, only the cross-correlation
coeﬃcient for lag -4 was signiﬁcant. This means that Trump’s tweets about
minorities were predicted by Clinton’s tweets about this topic four days earlier.
Table 2 also shows that for the topic terrorism, cross-correlation coeﬃcients
for several lags and leads were signiﬁcant (lag -3 and -1, lead 2 and 4). These
results imply that in contrast to the topic minorities, correlations between both
candidates’ time series are bi-directional. In other words, Trump’s tweets about
terrorism were predicted by Clinton’s tweets about this topic one and three
days earlier, but Clinton’s tweets are also predicted by Trump’s tweets two and
four days earlier. For the other topics (family and employment ), no signiﬁcant
cross-correlation coeﬃcients were found.
Table 1. Topic distribution of Donald Trump’s and Hillary Clinton’s tweets in descend-
ing order of frequency. Categories marked with * were the most frequently discussed
political topics and used in further analysis.
Category No. of cases % of cases Category No. of cases % of cases
1 Parties and
1783 58.36 Parties and
2 Media 778 25.47 Family* 488 16.04
3 Slogans 569 18.63 Media 446 14.66
493 16.14 Slogans 432 14.20
5 Caucus 296 9.69 Meetings 195 6.41
6 Employment* 129 4.22 Minorities* 189 6.21
7 Family* 92 3.01 Employment* 177 5.82
8 Terrorism* 83 2.72 Caucus 138 4.53
9 Trump family 78 2.55 Human rights 129 4.24
10 Minorities* 56 1.83 Healthcare 119 3.91
11 Healthcare 21 0.69 Thank-you
20 0.65 Environment 87 2.86
13 0.43 Guns 85 2.79
14 Human rights 11 0.36 Terrorism* 68 2.23
15 Justice 8 0.26 Education 52 1.71
7 0.23 Justice 49 1.61
17 Education 7 0.23 Economics 29 0.92
18 Economics 4 0.13 Military
19 Guns 4 0.13 Campaign
20 Environment 2 0.07 Foreign
Table 2. Cross-correlation results for the topics minorities, terrorism, employment and
Political topic Signiﬁcant lags and leads for
Trump with Clinton
Minorities Lag (-4) 0.20
Lag (-3) 0.16
Lag (-1) 0.32
Lead (2) 0.19
Lead (4) 0.18
Employment No signiﬁcant lags or leads
Family No signiﬁcant lags or leads
In Figs. 1 and 2, visualizations of the minorities and terrorism time series are
shown, as cross-correlation returned signiﬁcant results for these topics. It stands
out that both candidates tweet much less continuously about terrorism. Most of
the time numbers of daily tweets are quite low, but four diﬀerent peaks can be
identiﬁed during the tracking timespan. Fig. 2 shows that terror attacks took
place right before the ﬁrst three peaks. Since many signiﬁcant auto-correlations
were found in Trump’s terrorism time series and terror attacks took place right
before the peaks, the signiﬁcant cross-correlation should be interpreted carefully.
It is possible that these external events instead of agenda setting explain the
correlations between Trump’s and Clinton’s time series for the topic terrorism.
For the topic minorities no such external events were found, so agenda setting
might be a reasonable explanation for the signiﬁcant cross-correlation.
Number of tweets by Trump Nu mbe r of tweet s by Clinton
Fig. 1. Visualization of both candidates’ time series for the topic minorities.
Table 3 summarizes the results of the hypothesis tests. The ﬁndings provide
partial support for the ﬁrst hypothesis, and no support for the second hypothesis.
Table 3. Results of hypothesis tests
H1 The issues mentioned in Hillary Clinton’s tweets
will predict the issues mentioned in Donald
Supported for topic minorities
H2 The issues mentioned in Donald Trump’s tweets
will predict the issues mentioned in Hillary Clin-
Number of tweets by Trump Nu mbe r of tweet s by Clinton
Nov 13, 201 5: Terrorist atta ckin Paris, France
Dec 2, 201 5: Terrorist atta ck in San Berna dino, US
Mar 22, 201 6: Terrorist atta ck in Brussels, Belgium
Fig. 2. Visualization of both candidates’ time series for the topic terrorism.
This section discusses the results and addresses the research questions. It also
addresses the limitations of the study and mentions implications for future re-
5.1 Communication topics
The ﬁrst goal of this research was to study the nature of communication by
the eventual presidential nominees on Twitter. We found that political issues on
Twitter are not as frequent as thank-you messages and announcements for up-
coming media presences of the candidates. This conﬁrms the ﬁnding of Sandoval
et al. .
This ﬁnding also has some practical implications. Twitter serves as a source of
political information for many people , so politicians could likely beneﬁt from
disseminating more political statements on Twitter than they currently do. On
the other hand, Twitter only allows messages with up to 140 characters, which
may make it diﬃcult to discuss complex topics. Secondly, during his campaign,
Donald Trump was known for his plan of building a wall between Mexico and the
US to stop crime. He was also frequently mentioned in the media for his plans
to institute a law that would make the immigration of Muslims into the US
illegal . It can therefore be considered somewhat surprising that the number
of tweets from Donald Trump about minorities was very low. There were only
56 tweets regarding this issue in the tracking period. An explanation for this
divergence might be inter-media agenda setting. Perhaps, the statements were
so arousing or polarizing that journalists picked up on them and wrote or talked
about them much more than Donald Trump did himself.
A comparison of the candidates regarding speciﬁc topics reveals that both
politicians took them up diﬀerently. For example, after the terrorist attack of San
Bernadino, the biggest terror attack in the US , Donald Trump criticized that
it was not reported as a terrorist act (December 4, 2015). Hillary Clinton claimed
that the attack could only take place because “Republican Senators blocked a
bill to stop suspected terrorists from buying guns” (December 4, 2015). Here,
Hillary Clinton uses the terror attack to criticize the Republican Party. Donald
Trump also blamed Hillary Clinton indirectly in March, during the Brussels ter-
ror attacks: “Hillary Clinton has been working on solving the terrorism problem
for years. TIME FOR CHANGE, I WILL SOLVE - AND FAST” (March 24,
2016). During these bombings, Clinton twittered: “These terrorists seek to un-
dermine the democratic values that are the foundation of our way of life. They
will never succeed. -H.” (March 22, 2016). All in all, both candidates used ter-
rorist attacks for their purposes. When comparing the issue family, it becomes
clear that the perception that Hillary Clinton plays the “woman card”  may
be rooted in the fact that she addresses this topic ﬁve times as often as Trump.
5.2 Agenda setting
In addressing the second research question, this paper is the ﬁrst one to exam-
ine inter-policy agenda setting eﬀects on the microblogging service Twitter for
candidates during US primary elections. The goal was to answer the question
whether agenda setting takes place between the candidates during the US pres-
idential elections. This question can only partially be answered with yes. For
the topic minorities, Clinton is leading Trump’s agenda. Results for employment
and family were not signiﬁcant. We conclude that the presence of agenda setting
depends on the political issue. With this research, we contribute to the work
of Soroka , who introduced the notion of inter-policy agenda setting and
conﬁrmed existing research by Vliegenthart et al. .
For the topic terrorism, agenda setting might have taken place in both direc-
tions, but signiﬁcant auto-correlations were present even after removing linear
trends, and the cross-correlations observed may therefore be spurious. Fig. 2
shows that attacks took place right before the peaks, which may explain this
phenomenon. We therefore refrain from drawing further conclusions from the
A closer inspection of the tweets reveals that when Donald Trump takes up
the issues of other candidates and politicians, he frequently does so to attack
them. For example, Trump twittered on December 7, 2015: “Obama said in his
speech that Muslims are our sports heroes. What sport is he talking about, and
who? Is Obama proﬁling?” In this tweet, he takes up Obama’s speech to ask
questions, encouraging his community to think about it.
In summary, this study oﬀers some evidence that agenda setting actually
takes place on Twitter, but less than could have been expected. In some cases
(e.g. terrorism) tweets looked more likely to be prompted by external events.
This ﬁnding is important because it emphasizes further how little interaction
with the political opponent occurred on Twitter in this case. Thereby our re-
search contributes to the small existing body of literature on agenda setting in
the context of Twitter.
As any research, ours comes with limitations. Spanish tweets were not considered
in this analysis. Expanding the dictionaries with Spanish words would have been
very time-consuming. However, since Clinton sometimes tweeted in Spanish, this
decision could skew the results on the topic of minorities.
Furthermore, when creating a dictionary for content analysis, there are al-
ways ambiguities regarding the appropriate category for a particular word, and
building a dictionary is a process that involves subjective decisions. We made
use of the Keyword-in-Context tool to resolve these uncertainties and evaluated
the dictionary by calculating recall and precision for every category. While val-
idation results should be interpreted carefully, they suggest that the dictionary
was suitable for our purposes.
We found signiﬁcant auto-correlations in some time series, which led us to
exclude the topic of terrorism. Linear trends were controlled for by using de-
trended versions of the aﬀected time series as described in Conway et al. ,
but more sophisticated methods are available to remove auto-correlations.
Finally, as already mentioned, the results of the US primaries in 2016 were
highly unexpected. Almost no one thought that Donald Trump will be the candi-
date for the Republican Party. This provided a unique research setting, but it is
uncertain if this research is replicable for other contexts and for other countries.
5.4 Future Research
We propose further research to examine if the results can be generalized. Ad-
ditionally, given the highly emotional content of tweets, it seems worthwile to
examine if emotion has an inﬂuence on agenda setting on social media, espe-
cially on Twitter. For example, the diﬀerence between candidates regarding the
sentiment of their tweets could be examined in future research. It is also unclear
whether sentiment has an inﬂuence on agenda setting on Twitter.
While executing our time series analysis, we found auto-correlations in Trump’s
series for terrorism. Possible external events are predicting the time series of
Trump. In further research, a more complex model could be developed that re-
moves these auto-correlations or explicitly takes the inﬂuence of outside events
into account. Another valuable research direction would be the enhancement of
automatic classiﬁcation methods which allow the identiﬁcation of political topics
in election-related tweets even though topics change between elections.
Future research should also include tweets by the public to evaluate how
agenda setting takes place between politicians and citizens on twitter. However,
in this regard it has to be considered that Twitter is only used by a small
percentage of the whole population and mostly by younger people. Once again,
the inﬂuence of sentiment should be considered in this context.
In his 1996 analysis of how a new technology had reshaped political campaigning
in Texas, Jonathan Coopersmith stated that “the spread of modern information
technologies has greatly altered the face of politics” [14, p. 37]. The technology he
examined was the fax machine. Twenty years later, many of the observations he
made are equally true for Twitter: a ﬂood of data is being generated, information
can be disseminated rapidly, and there is consequently increased pressure on
political campaigns to make use of these new technologies eﬀectively.
But Twitter does not only enable campaigns to spread information rapidly,
it also allows researchers an unprecedented glimpse at the daily activities of
campaigns. When Wattal et al.  laid out their research agenda, they called
on researchers to examine how the political system might change as a result of
the Internet. They ask, “how might the web be used to support increased mutual
understanding and tolerance in political discourse”?
Twitter has indeed become one of the most important social media used by
campaigns – but our analysis showed that very little political discourse actually
takes place there. The medium is dominated by thank-you messages and simple
political slogans. Candidates use it to reach out to their followers, not to engage
with the political opposition.
Still, the few political messages present in the data open new avenues for
researchers. Previous research on agenda setting had considered the agenda of
political actors and institutions in power. Now, a large part of the communication
by political campaigns is readily available to researchers in a digitized form. We
can analyze the agenda of those who will wield political power even before they
do it. Twitter has made it possible to carry out this analysis with less eﬀort and
at a larger scale than before.
In this study, we combined content analysis and time series analysis. Through
the resulting analysis of day-to-day frequencies of topic mentions, we were able
to ﬁnd little evidence for inter-policy agenda setting during the run-up to the
US presidential elections. Instead of fostering discussion and helping mutual
understanding, Twitter seems to represent a fractured social space.
1. Sabato, L.J., Kondik, K., Skelley, G.: Republi-
cans 2016: What To Do With The Donald? (2015),
2. Lee, J.C., Quealy, K.: The 337 People, Places and Things Don-
ald Trump Has Insulted on Twitter: A Complete List (2016),
3. Sperling, V.: Masculinity, Misogyny, and Presidential Image-making in
the U.S. and Russia (2016), https://global.oup.com/academic/category/social-
4. Button, M.E.: Trump and the Triumph of Hubris over Democratic Politics
5. Schwartzman, P., Johnson, J.: It’s not chaos. It’s Trump’s cam-
paign strategy (2015), https://www.washingtonpost.com/politics/its-not-
6. Cossu, J.V., Dugue, N., Labatut, V.: Detecting Real-World Inﬂuence through Twit-
ter. In: ENIC Proceedings. pp. 83–90 (2015)
7. Stieglitz, S., Dang-Xuan, L., Bruns, A., Neuberger, C.: Social Media Analytics –
An Interdisciplinary Approach and Its Implications for Information Systems. Bus.
Inf. Syst. Eng. 56(2), 101–109 (2014)
8. Cazzoli, L., Sharma, R., Treccani, M., Lillo, F.: A Large Scale Study to Understand
the Relation between Twitter and Financial Market. In: ENIC Proceedings. pp.
9. Dang-Xuan, L., Stieglitz, S., Wladarsch, J., Neuberger, C.: An Investigation of
Inﬂuentials and the Role of Sentiment in Political Communication on Twitter
During Election Periods. Inf. Commun. Soc. 16(5), 795–825 (2013)
10. Larsson, A.O., Moe, H.: Studying political microblogging: Twitter users in the
2010 Swedish election campaign. New Media Soc. 14(5), 729–747 (2012)
11. Sandoval, R., Matus, R.T., Rogel, R.N.: Twitter in Mexican Politics: Messages to
People or Candidates? In: AMCIS Proceedings. pp. 1–10 (2012)
12. Wattal, S., Schuﬀ, D., Mandviwalla, M., Williams, C.B.: Web 2.0 and Politics: The
2008 US Presidential Election and An E-Politics Research Agenda. MIS Q. 34(4),
13. Maldonado, M., Sierra, V.: Can Social Media Predict Voter Intention in Elections?
The Case of the 2012 Dominican Republic Presidential Election. In: AMCIS Pro-
14. Coopersmith, J.: Texas Politics and the Fax Revolution. Inf. Syst. Res. 7(1), 37–51
15. B¨uhler, J., Bick, M.: The impact of social media appearances during election cam-
paigns. In: AMCIS Proceedings. vol. 5, pp. 3406–3416 (2013)
16. Conway, B.A., Kenski, K., Wang, D.: The Rise of Twitter in the Political Cam-
paign: Searching for Intermedia Agenda-Setting Eﬀects in the Presidential Primary.
J. Comput. Mediat. Commun. 20(4), 363–380 (2015)
17. Gruszczynski, M.W.: Examining the Role of Aﬀective Language in Predicting the
Agenda-Setting Eﬀect. In: APSA Annual Meeting (2011)
18. Baran, S.J., Davis, D.K.: Mass Communication Theory: Foundations, Ferment,
and Future, vol. 6. Wadsworth, Boston (2015)
19. Cohen, B.C.: The Press and Foreign Policy. Princeton University Press, Princeton,
20. Dearing, J.W., Rogers, E.M.: Communication Concepts 6. Agenda-Setting. Sage,
Thousand Oaks, CA (1996)
21. Soroka, S.N.: Issue Attributes and Agenda-Setting by Media, the Public, and Pol-
icymakers in Canada. Int. J. Public Opin. Res. 14(3), 264–285 (2002)
22. Krasa, S., Polborn, M.: The binary policy model. J Econ. Theory 145(2), 661–688
23. Colomer, J.M., Llavador, H.: An agenda-setting model of electoral competition.
SERIEs 3(1-2), 73–93 (2012)
24. Baumgartner, F.R., Green-Pedersen, C., Jones, B.D.: Comparative Studies of Pol-
icy Agendas. J. Eur. Public Policy 13(7), 959–974 (2006)
25. Wolfe, M.: Putting on the brakes or pressing on the gas? Media attention and the
speed of policymaking. Policy Stud. J. 40(1), 109–126 (2012)
26. Peake, J.S.: Presidential Agenda Setting in Foreign Policy. Polit. Res. Q. 54(1),
27. Eshbaugh-Soha, M.: Presidential Inﬂuence of the News Media: The Case of the
Press Conference. Polit. Commun. 30(4), 548–564 (2013)
28. Rutledge, P.E., Larsen Price, H.A.: The President as Agenda Setter-in-Chief: The
Dynamics of Congressional and Presidential Agenda Setting. Policy Stud. J. 42(3),
29. Kiousis, S., Shields, A.: Intercandidate agenda-setting in presidential elections:
Issue and attribute agendas in the 2004 campaign. Public Relat. Rev. 34(4), 325–
30. Statista: Number of monthly active Twitter users in the United States
from 1st quarter 2010 to 2nd quarter 2016 (in millions) (2016),
31. Vargo, C.J.: Twitter as Public Salience: An Agenda-Setting Analysis. In: AEJMC
Annual Conference (2011)
32. Meraz, S.: Is There an Elite Hold? Traditional Media to Social Media Agenda
Setting Inﬂuence in Blog Networks. J. Comput. Mediat. Commun. 14(3), 682–707
33. Vergeer, M.: Twitter and Political Campaigning. Sociol. Compass 9(9), 745–760
34. Robertson, S.P., Vatrapu, R.K., Medina, R.: Oﬀ the wall political discourse: Face-
book use in the 2008 U.S. presidential election. Inf. Polity 15(1-2), 11–31 (2010)
35. Groshek, J., Clough Groshek, M.: Agenda Trending: Reciprocity and the Predictive
Capacity of Social Networking Sites in Intermedia Agenda Setting across Topics
over Time. Media Commun. 1(1), 15–27 (2013)
36. Vargo, C.J., Guo, L., Mccombs, M., Shaw, D.L.: Network Issue Agendas on Twitter
During the 2012 U.S. Presidential Election. J. Commun. 64(2), 296–316 (2014)
37. Ancu, M., Cozma, R.: MySpace Politics: Uses and Gratiﬁcations of Befriending
Candidates. J. Broadcast. Electron. Media 53(4), 567–583 (2009)
38. Vliegenthart, R., Walgrave, S., Meppelink, C.: Inter-party Agenda-Setting in the
Belgian Parliament: The Role of Party Characteristics and Competition. Polit.
Stud. 59(2), 368–388 (2011)
39. Tedesco, J.C.: Intercandidate Agenda Setting in the 2004 Democratic Presidential
Primary. Am. Behav. Sci. 49(1), 92–113 (2005)
40. Liu, B., Hu, M., Cheng, J.: Opinion Observer: Analyzing and Comparing Opinions
on the Web. In: WWW Proceedings. pp. 342–351 (2005)
41. CNN: San Bernadino Shooting (2016), http://edition.cnn.com/specials/san-
42. Zezima, K.: Trump: Clinton is playing the ’woman card’ (2016),